“A small Sokoban-like puzzle: a lighthouse keeper pushes oil drums onto the lamp pads to relight the lighthouses. Five short levels that get harder.”
A separate Claude Code session was the Game Director for an empty game, on the grid-puzzle template. It had Gamewright's tools and nothing else: no shell, no way to read files, no way to write code. Everything it learned came through the tools, and everything it changed was a proposal. The person decided on each one, answered briefly, and looked at the game before saying it was played.
The person’s words are quoted as they were said. The rest is what the Director did, in plain words, from the run’s own record.
“A small Sokoban-like puzzle: a lighthouse keeper pushes oil drums onto the lamp pads to relight the lighthouses. Five short levels that get harder.”
Wrote a brief (no one gets hurt; undo and restart always work), the game, and a first playable version: the keeper, drums, pads, rocks, the push rule and five levels.
Each level teaches one idea: a straight push, getting behind a drum, a dead-end shaft, three pads close together, and an order in which to light the lamps.
The template has two questions. It answered them from the description rather than guess silently, and said so.
The brief, the game and the first playable version.
The solver proved all five levels can be won, and a playtest replayed each solution in the engine.
Played it, then said:
“mark it played, then a rocky lighthouse coast at dusk, drums and pads that read at a glance, a lamp that lights when a pad is filled”
Marked the game as played. Then the look and a new rule: a drum on a pad becomes a lit lamp, and goes back to a drum if pushed off; a level is won when every pad is lit.
It put an islet in a sea around every level, and started the first level, First Light, with one lamp already lit, so the goal shows.
Nothing it knew of could colour the grey margins around the board, so it asked the engine for that.
Four proposals: dusk colours, the lit-lamp rule, the islet, and a first lamp lit.
With the new rule, the solver proved all five levels again.
“The margins still read as an editor”
Lost its tools both times: Claude Code gave up on Gamewright's server while listing the game's documents (see what the run found, below).
Said the same again, after the server was fixed.
Searched everything the engine can do, and said plainly that nothing could colour the margins. It named the nearest miss, a gradient drawn over the top of the game, and advised against it.
“The engine team built your backdrop request”
Found the new colour slots on its own, and made the sky in one change that touched only the look: indigo overhead to a coral glow at the foot.
No check measured text contrast yet, so it worked that out by hand. And it said what it hadn't seen: levels 2 to 5, and the panel shown when a level is solved.
One change to the look: the sky.
Lamplighter: push every oil drum onto a lamp pad. A drum on a pad becomes a lit lamp, and a level is won when every lamp burns. Five levels, each teaching one idea, on an islet under a dusk sky.
Plays on a phone or tablet too, with on-screen controls.
Each finding as the run’s write-up recorded it, and what has happened since.
Claude Code lists what a server offers as it connects, and gives up after 30 seconds. Listing the game's documents ran the whole game check, which solves every level: 29 seconds on Lamplighter.
Listing and reading a document now only reads the game, and the solver remembers the levels it has solved. The server now answers in 1.5 s, and turn 6 took 229 s where turn 2 took 2,376 s.
The grid view read one backdrop colour, named only inside a sentence of its description.
Views now declare the colours they read, the vocabulary lists them, and the grid view takes a top and a bottom colour for a sky. The Director found them on its own on its next turn.
The Director worked it out by hand.
Being added when the run was written up. Since then, the frame check measures the contrast of every text the engine draws, and fails it under 3:1.